Axel Vuorinen 1, Emmanuelle Comets 2,3, Moreno Ursino 1
1 Inserm, Université Paris Cité, Inria, HeKA (F-75015 Paris, France), 2 Univ Rennes, Inserm, EHESP, Irset (Institut de recherche en santé, environnement et travail) - UMRS 1085 (F-35000 Rennes, France), 3 Université Paris Cité and Université Sorbonne Paris Nord, Inserm, IAME (F-75018 Paris, France)
Introduction: Early-phase dose-finding generally focuses on toxicity, but new approaches are shifting to a more comprehensive quantification of benefit-risks including efficacy assessment[1]. These approaches leverage all available data sources to inform trial design, including population PK/PD modelling and exposure–activity–response analyses, with the goal of avoiding the advancement of suboptimal regimens into later-phase trials [2,3]. Incorporating PK/PD measures into the modelling framework allows regimens to be compared based on actual drug exposure and biological effect, accounts for interpatient variability, and captures differences in how dosing schedules influence drug levels in the body. An adaptive platform trial (APT) uses a single master protocol with a shared control, interim analyses, and prespecified (often Bayesian) decision rules to continuously evaluate multiple treatments, allowing arms to be adaptively added, dropped, or preferentially allocated based on accumulating patient data. Our objective was to develop a phase II APT design, to compare dosing regimens proposed for an investigational drug, that integrates drug exposure and pharmacologic activity, estimated through a joint population PK/PD model, into a Bayesian decision framework.
Methods: We propose a two-stage framework for an APT where each experimental arm evaluates a distinct dosing regimen. This APT incorporates joint population PK/PD modelling to link dosing regimens to drug exposure and biological activity, enabling patient- and population-level predictions that inform toxicity and efficacy assessments. In the first stage, patients are allocated to the arm only based on toxicity outcomes. In the second stage, starting with the first interim analysis, patients are allocated using the activity and efficacy. The proposed two-stage framework. Building on the DICE and TITE-PK models [4,5], we propose a time-to-event (TTE) model for acute and cumulative toxicities using pharmacokinetics (TACT-PK), in which follow-up is discretised into dosing cycles. This enables estimation of per-cycle toxicity risk, accommodates irregular event timing, and allows incorporation of exposure metrics. No ordering assumption among regimens is required. This model allows early identification and discontinuation of unsafe regimens. A Bayesian model-averaged TTE approach, involving two complementary TTE models, one linking efficacy to a longitudinal PD biomarker and another relating it directly to cumulative exposure when the biomarker is weakly predictive, enables robust regimen optimisation and assessment of the added value of activity data.
The design is evaluated in an influenza ICU setting, under the assumption that a prior dose-finding study has identified three dosing regimens. Drug concentration is described by a one-compartment model with first-order absorption and linear elimination. The activity biomarker is modelled using a semi-mechanistic cytokine turnover model with drug-dependent inhibition through a sigmoidal Imax function, together with a time-varying negative feedback capturing immune refractoriness and adaptive attenuation of the IL6-response [6,7]. Efficacy is defined as time to ICU discharge.
Results: The proposed design was evaluated through a simulation study and compared with a benchmark dose-based platform design using Bayesian models for toxicity (DICE [4]) and efficacy (Bayesian Cox model [8]). Performance was assessed across key operating characteristics, including correct graduation of the optimal regimen, time to graduation, estimation accuracy, and patient allocation. The three investigational regimens evaluated against a common control all had higher true efficacy than the control, although only one represented the optimal benefit–risk profile. Both designs achieved a similar probability of correctly graduating the optimal regimen (approximately 61%). The PK/PD-informed design demonstrated greater efficiency, with earlier arm graduation on average (10.7 vs 12.1 interim analyses). It also produced more accurate toxicity estimates overall and a better-calibrated assessment of benefit–risk, particularly by capturing cumulative toxicity with the TACT-PK model. In contrast, the dose-based design tended to overestimate efficacy and underestimate early toxicity for higher-risk regimens. Patient allocation patterns were broadly similar across designs.
Conclusion: Integrating PK/PD modelling into the adaptive platform design improves the ability to discriminate among dosing regimens based on the exposure–response relationships. Future investigations will involve the incorporation of additional models reflecting other biological and clinical mechanisms, as well as the use of covariates to identify patient subgroups for whom the therapy is most beneficial.
References:
[1] FDA (2021). Oncology Center of Excellence: Project Optimus. https://www.fda.gov/about-fda/oncology-center-excellence/project-optimus
[2] Fourie Zirkelbach J et al. J Clin Oncol, 40:3489–3500, 2022.
[3] Gao W et al. CPT:PSP, 13(5):691–709, 2024.
[4] Ursino M, Biard L, Chevret S. Biom J, 64:1486–1497, 2022.
[5] Günhan BK, Weber S, Friede T. Stat Med, 39:3986–4000, 2020.
[6] Chen X et al. Clin Transl Sci, 12:600–608, 2019.
[7] Lefèvre A et al. Front Immunol, 15:1463915, 2025.
[8] Jaki T et al. Stat Methods Med Res, 33(11-12):2115–2130, 2024.
Acknowledgments: This study was supported by a grant from Inserm and the French Ministry of Health (MESSIDORE 2022, reference number Inserm-MESSIDORE N° 94).
Reference: PAGE 34 (2026) Abstr 12259 [www.page-meeting.org/?abstract=12259]
Poster: Methodology - Study Design